Water-ice classification method and system for multispectral remote sensing image based on double-knowledge distillation

Abstract

The invention belongs to the technical field of remote sensing image water-ice classification, and particularly relates to a water-ice classification method and system for a multispectral remote sensing image based on double-knowledge distillation. The method comprises the following steps: step 1, acquiring a multispectral remote sensing image and implementing data preprocessing; step 2, constructing a glacial lake extraction data-free knowledge distillation network, adopting a glacial lake index ILI as a first teacher model, constructing a U-net network as a first student model, and completing data-free knowledge distillation; step 3, constructing a glacial lake extraction data-knowledge-free distillation framework; 4, constructing a water-ice classification teacher-student off-line distillation network which is used for performing water-ice classification on the image after the glacial lake is extracted; step 5, constructing a water-ice classification knowledge distillation framework; and 6, taking a remote sensing image as input, carrying out glacial lake extraction through the glacial lake extraction data-knowledge-free distillation framework in the step 3, and carrying out water-ice classification operation through the water-ice classification teacher-student off-line distillation framework in the step 5.

Description

Technical Field

[0001] The present invention belongs to the technical field of water ice classification in remote sensing images, and in particular relates to a water ice classification method and system for multispectral remote sensing images based on dual knowledge distillation. Background Art

[0002] With the advancement of remote sensing technology and machine learning, water ice classification research has received increasing attention worldwide, especially in the monitoring of ice and snow cover in polar regions such as the Arctic and Antarctic, where remarkable results have been achieved. Many scientific research institutions have conducted long-term monitoring of glaciers, ice sheets, and lake ice based on high-resolution satellite data (such as Landsat, MODIS, and Sentinel), using traditional remote sensing classification methods (such as supervised classification and unsupervised classification) and deep learning methods to segment ice and snow areas and analyze their dynamic changes.

[0003] In recent years, deep learning has made significant breakthroughs in image processing. Convolutional neural networks (CNNs) have been gradually applied to water ice classification tasks, greatly improving classification accuracy. For example, by combining multispectral or multi-temporal remote sensing images, CNN-based automated water ice classification research has greatly improved the ability to identify different types of water ice. Despite this, water ice classification still faces many challenges. Remote sensing images usually contain ground information other than water ice (such as land), which can easily introduce noise and interfere with classification accuracy. To solve this problem, in the existing literature "Icy lakes extraction and water-ice classification using Landsat 8 OLI multispectral data", a water ice classification method combining index analysis and classification is proposed: first, the ice lake identification index (ILI) is calculated using the bands of Landsat 8 OLI multispectral images, thereby extracting the ice lake area, and then classifying the water ice. This method effectively isolates the interference area and improves the accuracy of classification.

[0004] At present, research in the field of water ice classification mainly focuses on using deep learning models such as convolutional neural networks (CNN) and residual neural networks (ResNet) to distinguish between new ice, one-year ice, multi-year ice, and open water. However, as the complexity of the model structure increases, the convergence time of the training process is significantly prolonged, and it becomes more difficult to deploy on devices with limited resources, which hinders the promotion of deep learning models in practical applications.

[0005] In summary, water ice classification tasks face challenges such as interference from ground objects and lack of labels. Complex deep learning models also have high demands on hardware resources, which limits their application potential. Therefore, this field urgently needs to seek new breakthroughs in model design and algorithm optimization to achieve more efficient and reliable water ice classification technology. Summary of the invention

[0006] In view of the defects that the manual annotation of label values ​​in the current deep learning-based remote sensing image water ice classification task is time-consuming and labor-intensive, and water ice classification cannot be effectively performed under complex conditions, the present invention provides a water ice classification method and system for multispectral remote sensing images based on dual knowledge distillation, which can combine multispectral and multi-temporal data to improve the accuracy and robustness of water ice classification, and can effectively train the model in the absence of annotated data sets, thereby improving the operating efficiency and accuracy of the model.

[0007] The present invention adopts the following technical scheme: The water ice classification method of multispectral remote sensing images based on dual knowledge distillation includes the following steps: Step 1, acquiring multispectral remote sensing images and performing data preprocessing; Step 2: Build the Ice Lake Extraction Dataless Knowledge Distillation Network, take the Ice Lake Index ILI as the first teacher model, build the U-net network as the first student model, and complete the dataless knowledge distillation; Step 3, constructing an Ice Lake Extraction Dataless Knowledge Distillation Framework, inputting the data preprocessed in Step 1 as a training set into the Ice Lake Extraction Dataless Knowledge Distillation Network described in Step 2, outputting the binary masks of Ice Lake Extraction respectively, and performing DICE loss calculation on the Ice Lake mask output of the first teacher model ILI and the Ice Lake mask output of the first student model U-net as the supervision loss; Step 4: construct a water ice classification teacher-student offline distillation network to classify water ice in images after ice lake extraction; the second teacher model uses a trained U-net classification model, and the second student model uses a lightweight model; Step 5, construct a knowledge distillation framework for water ice classification, perform ice lake extraction on the image through the first student model trained in step 3, and use the output of the first student model as the input of the water ice classification teacher-student offline distillation network in step 4 to train water ice classification; wherein the second teacher model and the second student model in the water ice classification teacher-student offline distillation network respectively output water ice classification results; In step 6, the remote sensing image is used as input, and ice lake extraction is performed through the ice lake extraction data-free knowledge distillation framework in step 3, and then water ice classification is performed through the water ice classification teacher-student offline distillation framework in step 5.

[0008] Preferably, step 1 comprises the following steps: S11, grouping the multispectral remote sensing images by band according to spatial resolution, and reconstructing the multispectral remote sensing images with different spatial resolutions by super-resolution, and uniformly processing them to the highest spatial resolution; S12, cutting the image reconstructed in step S11 to form a training set.

[0009] Preferably, in step 2, the U-net network includes an encoder, a decoder and an output layer; and the number of parameters of the first student model is smaller than the number of parameters of the first teacher model.

[0010] Preferably, in step 3, the supervision loss is used as a network training error to update and iterate the first student model, so that the first student model can learn the feature knowledge of the first teacher model to the greatest extent, thereby achieving data-free knowledge distillation; the supervision loss is calculated using the DICE function and used for back propagation to guide the training of the first student model, which is specifically expressed as: ; in, is the loss function of the Ice Lake extraction network, M represents the total number of pixels in the image, is the sum of the probabilities of the predicted segmentation results at pixel i; is the binary value of the true label at pixel i.

[0011] Preferably, in step 4, the second teacher model includes an encoder, a decoder and an output layer; the second student model has the same architecture as the second teacher model, and the number of parameters of the second student model is smaller than the number of parameters of the second teacher model.

[0012] Preferably, step 5 further comprises the following steps: S51, through Fourier transformation, the domain-invariant knowledge and domain-specific knowledge of the second teacher model are separated. When training the teacher-student model, the second student model acquires the domain-invariant knowledge by learning the second teacher model and acquires the domain-specific knowledge through labels; S52, the output of the second teacher model is compared with the output of the second student model as the teacher supervision loss, and the data supervision is performed by manually annotating the generated label data and comparing it with the second student model as the data supervision loss; The teacher supervision loss and data supervision loss are both calculated using the DICE function and used for back propagation to guide the training of the second student model, specifically expressed as: ; ; ; in, is the loss function for water ice classification, is the data supervision loss, is the teacher supervision loss; N represents the total number of pixels in the image; is the sum of the probabilities of the second student model predicting the segmentation result at pixel i; is the sum of the probabilities of the second teacher model predicting the segmentation result at pixel i, is the binary value of the true label segmentation result at pixel i; and They are weight coefficients for balancing losses in different aspects.

[0013] The present invention also discloses a water ice classification system for multispectral remote sensing images based on dual knowledge distillation, based on the above method, including the following modules: Image acquisition and preprocessing module, which acquires multispectral remote sensing images and performs data preprocessing; Ice Lake extracts data-free teacher-student offline distillation network construction modules, builds a teacher-student offline distillation network based on multispectral remote sensing images, takes the Ice Lake Index ILI as the first teacher model, and builds a U-net network as the first student model; The loss calculation module builds a knowledge distillation framework for Ice Lake Extraction for multispectral remote sensing images, inputs the training set in the image acquisition and preprocessing module into the first teacher model and the first student model in the Ice Lake Extraction Dataless Teacher-Student Offline Distillation Network Construction Module, outputs the Ice Lake mask respectively, and calculates the loss of the outputs of the first teacher model and the first student model as the supervision loss; The water ice classification teacher-student offline distillation network construction module builds a water ice classification teacher-student offline distillation network based on multispectral remote sensing images, classifies water ice in images after ice lake extraction, uses the trained U-net classification model as the second teacher model, and uses the lightweight model as the second student model; The total loss module constructs a water ice classification knowledge distillation framework for multispectral remote sensing images, performs ice lake extraction on the image through the first student model in the trained ice lake extraction data-free knowledge distillation framework, and uses the output of the first student model as the input of the water ice classification teacher-student offline distillation network to train water ice classification; the teacher model and the student model of the water ice classification teacher-student offline distillation network output water ice classification results respectively, and the outputs of the second teacher model and the second student model are calculated as teacher supervision loss, i.e., knowledge distillation loss; the super-resolution remote sensing image output by the second student model and the training set label are calculated as data supervision loss; the knowledge distillation loss and the data supervision loss are added by weight to calculate the total loss; The water ice classification module takes the multispectral remote sensing image as input, obtains the ice lake mask and extracts the ice lake through the first student model trained in the ice lake extraction student network training module, and then inputs the extracted ice lake image into the second student model trained in the water ice classification student network training module to obtain the water ice classification result.

[0014] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention uses knowledge distillation to achieve network training and deployment for effective extraction of ice lakes without data supervision and with a simple model structure, solving the problem of extracting ice lakes when there are few true values ​​and it is more complicated to make true labels; (2) The present invention uses domain theory in knowledge distillation to effectively transfer domain-invariant knowledge applicable to the student model and discard domain-specific knowledge. This solves the problem of using domain theory to more effectively transfer information shared by different networks when a large network model of the target environment is missing. (3) The double-layer knowledge distillation method proposed in the present invention can not only improve the training speed of the lightweight student network, but also enable the student network to achieve the effect of the teacher network with a large number of parameters under the premise of a small number of parameters. It can also perform more accurate and effective water ice classification after effectively eliminating interference, making it more suitable for deployment in embedded devices with limited resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of the entire process of network design-training-deployment of a water ice classification method for multispectral remote sensing images based on dual knowledge distillation provided in Example 1 of the present invention; Figure 2 A schematic diagram of a process for ice lake extraction training in a water ice classification method for multispectral remote sensing images based on dual knowledge distillation provided in Example 1 of the present invention; Figure 3 A schematic diagram of a process flow of water ice classification training in a water ice classification method for multispectral remote sensing images based on dual knowledge distillation provided in Example 1 of the present invention; Figure 4 A flow chart of a practical application of the water ice classification method for multispectral remote sensing images based on dual knowledge distillation provided in Example 1 of the present invention Figure 5 A principle block diagram of a water ice classification system for multispectral remote sensing images based on dual knowledge distillation provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0016] In order to more clearly illustrate the embodiments of the present invention, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other accompanying drawings and other implementation methods can be obtained based on these accompanying drawings without creative work.

[0017] Embodiment 1: This embodiment proposes a water ice classification method for multispectral remote sensing images based on dual knowledge distillation. Figure 1 This is a schematic diagram of the network design-training-deployment process of the water ice classification method for multispectral remote sensing images based on dual knowledge distillation. The remote sensing image is passed through the trained ice lake extraction network and the water ice classification network to obtain the water ice classification result. The specific steps are as follows: Step 1, acquiring multispectral remote sensing images and performing data preprocessing; Step 2: Build the Ice Lake Extraction Dataless Knowledge Distillation Network, take the Ice Lake Index ILI as the first teacher model, build the U-net network as the first student model, and complete the dataless knowledge distillation; the output of the first teacher model and the output of the first student model are lost through DICE; Step 3, constructing an Ice Lake Extraction Dataless Knowledge Distillation Framework, inputting the data preprocessed in Step 1 as a training set into the Ice Lake Extraction Dataless Knowledge Distillation Network described in Step 2, outputting the binary masks of Ice Lake Extraction respectively, and performing DICE loss calculation on the Ice Lake mask output of the first teacher model ILI and the Ice Lake mask output of the first student model U-net as the supervision loss; Step 4: construct a water ice classification teacher-student offline distillation network to classify water ice in images after ice lake extraction; the second teacher model uses a trained U-net classification model, and the second student model uses a lightweight model; Step 5, construct a knowledge distillation framework for water ice classification, perform ice lake extraction on the image through the first student model trained in step 3, and use the output of the first student model as the input of the water ice classification teacher-student offline distillation network in step 4 to train water ice classification; wherein the second teacher model and the second student model in the water ice classification teacher-student offline distillation network respectively output water ice classification results; In step 6, the remote sensing image is used as input, and ice lake extraction is performed through the ice lake extraction data-free knowledge distillation framework in step 3, and then water ice classification is performed through the water ice classification teacher-student offline distillation framework in step 5.

[0018] This embodiment takes the L2A level (the level after atmospheric correction processing of Sentinel-2) image of Sentinel-2 / MSI as an example, and the specific steps are as follows: Step 1: Log in to the website to download Sentinel-2 image data and perform preprocessing. Specifically, based on the characteristics of Sentinel-2 remote sensing images, adopt band grouping (10m, 20m, 60m GSD), and perform super-resolution reconstruction of 20m and 60m GSD to 10m band.

[0019] Step 2: Build the Ice Lake Extraction Dataless Knowledge Distillation Network, where the first teacher model uses ILI and builds the U-net network as the first student model. The structure of U-net is encoder, decoder and output layer to complete the dataless knowledge distillation. The output of the teacher model and the output of the student model are lost through dice.

[0020] Step 3: If Figure 2 As shown, a data-free knowledge distillation framework for Ice Lake extraction is constructed, the image training set described in step 1 is input into the first teacher model and the first student model described in step 2, and the Ice Lake mask is output respectively. The outputs of the two models are used for loss calculation as the teacher supervision loss (knowledge distillation loss), so as to prompt the student network to learn the features of the teacher network and the features of the Ice Lake mask label; Step 4: Construct a water ice classification teacher-student offline distillation network to classify water ice in images after extracting ice lakes. The second teacher model uses a trained U-net classification model, and the second student model uses a lightweight model with a smaller number of parameters. Step 5: Figure 3 As shown in the figure, a knowledge distillation framework for water ice classification is constructed, and the ice lake mask obtained in step 3 is processed and input into the water ice classification teacher-student network in step 4. The water ice images are output separately, and the loss of the two is calculated as the teacher supervision loss (knowledge distillation loss), which prompts the student network to learn the characteristics of the teacher network; the super-resolution remote sensing image output by the student network and the training set label are calculated as the data supervision loss, which prompts the student network to learn the characteristics of the high-resolution label. The teacher supervision loss and data supervision loss are added according to the weights to calculate the total loss.

[0021] Step 6: Figure 4 As shown in the figure, the Sentinel-2 image is used as input, and the ice lake extraction data-free knowledge distillation framework in step 3 is used to extract the ice lake from the Sentinel-2 image. Then, the image after ice lake extraction is used as input, and the water ice classification is completed through the teacher-student offline distillation network in step 5.

[0022] In step 1 of this embodiment, it is necessary to perform super-resolution processing on remote sensing images with different spatial resolutions and uniformly process them to the highest spatial resolution.

[0023] In step 2 of this embodiment, the data-free knowledge distillation network for ice lake extraction only needs the ice lake mask output by the first teacher model as supervision, without additional data supervision, thereby reducing consumed resources. The first teacher model is ILI, which is a specific implementation scheme for extracting ice lake.

[0024] In step 3 of this embodiment, the teacher supervision loss is calculated using DICE, and the teacher supervision loss is used as the network training error. After updating the iterative training, the final first student model is obtained, so that the student model can learn the characteristic knowledge of the teacher model to the greatest extent, thereby achieving data-free knowledge distillation. The loss function is DICE, and the back propagation guides the model training, which is specifically expressed as: ; in, is the loss function of the Ice Lake extraction network, M represents the total number of pixels in the image, is the sum of the probabilities of the predicted segmentation results at pixel i; is the binary value of the true label at pixel i. After obtaining the ice lake mask, the data needs to be further processed. The processing process is to combine the ice lake binary mask with the image data and set the pixel value of the image to 0 at the non-ice lake to eliminate the influence of other objects except the ice lake.

[0025] In step 4 of this embodiment, the pre-trained network U-net is used as the teacher network (second teacher model). The main components of the teacher network structure are the encoder, the decoder and the output layer. The student network (second student model) shares the same architecture as the teacher network (second teacher model), but has fewer parameters.

[0026] In step 5 of this embodiment, Figure 3 As shown in the figure, the domain-invariant knowledge and domain-specific knowledge of the second teacher model are separated through Fourier transformation. When training the teacher-student network, the second student model acquires domain-invariant knowledge by learning the second teacher model and acquires domain-specific knowledge through labels. The output of the second teacher model and the output of the second student model are used as teacher supervision loss, while data supervision generates label data through rapid manual annotation and uses loss with the second student model as data supervision loss. The loss function is DICE, and back propagation guides model training, which is specifically expressed as: ; ; ; in, is the loss function for water ice classification, is the data supervision loss, is the teacher supervision loss; N represents the total number of pixels in the image; is the sum of the probabilities of the second student model predicting the segmentation result at pixel i; is the sum of the probabilities of the second teacher model predicting the segmentation result at pixel i, is the binary value of the true label segmentation result at pixel i; and They are weight coefficients for balancing losses in different aspects.

[0027] The specific implementation steps of this embodiment include the following steps based on logging into the website to download Sentinel-2 image data and performing preprocessing: 1. Constructing a training set: Since Sentinel-2 images have three spatial resolutions of 10m, 20m, and 60m, the three different resolutions are first grouped, and the 20m spatial resolution and 60m spatial resolution are super-reconstructed to 10m spatial resolution; the network in step 2 directly inputs the super-resolution reconstructed Sentinel-2 image; the water ice classification network in step 4 needs to further segment the image.

[0028] 2. Design a lightweight student network: Both student networks (the first student model and the second student model) are based on the U-net network. While using the encoder and decoder, by adopting "skip connection", the network can extract very deep features.

[0029] 3. Construct loss function: The two teacher networks (the first teacher model and the second teacher model) are processed separately. The ice lake extraction teacher network is directly used to build the system, and the water ice classification teacher network loads pre-trained weights. Both teacher-student networks use DICE loss.

[0030] 4. Training model: The loss is taken as the training error of the classroom-student network, and the final student network model (first student model and second student model) is obtained after updating iterative training.

[0031] 5. Use the trained student network model (first student model and second student model) to perform ice lake extraction and water ice classification, effectively improving the efficiency of water ice classification.

[0032] Embodiment 2: like Figure 5 As shown, this embodiment discloses a water ice classification system for multispectral remote sensing images based on dual knowledge distillation, which is used to execute the method of embodiment 1, and includes the following modules: Image acquisition and preprocessing module, which acquires multispectral remote sensing images and performs data preprocessing; Ice Lake extracts data-free teacher-student offline distillation network construction modules, builds a teacher-student offline distillation network based on multispectral remote sensing images, takes the Ice Lake Index ILI as the first teacher model, and builds a U-net network as the first student model; The loss calculation module builds a knowledge distillation framework for Ice Lake Extraction for multispectral remote sensing images, inputs the training set in the image acquisition and preprocessing module into the first teacher model and the first student model in the Ice Lake Extraction Dataless Teacher-Student Offline Distillation Network Construction Module, outputs the Ice Lake mask respectively, and calculates the loss of the outputs of the first teacher model and the first student model as the supervision loss; The water ice classification teacher-student offline distillation network construction module builds a water ice classification teacher-student offline distillation network based on multispectral remote sensing images, classifies water ice in images after ice lake extraction, uses the trained U-net classification model as the second teacher model, and uses the lightweight model as the second student model; The total loss module constructs a water ice classification knowledge distillation framework for multispectral remote sensing images, performs ice lake extraction on the image through the first student model in the trained ice lake extraction data-free knowledge distillation framework, and uses the output of the first student model as the input of the water ice classification teacher-student offline distillation network to train water ice classification; the teacher model and the student model of the water ice classification teacher-student offline distillation network output water ice classification results respectively, and the outputs of the second teacher model and the second student model are calculated as teacher supervision loss, i.e., knowledge distillation loss; the super-resolution remote sensing image output by the second student model and the training set label are calculated as data supervision loss; the knowledge distillation loss and the data supervision loss are added by weight to calculate the total loss; The water ice classification module takes the multispectral remote sensing image as input, obtains the ice lake mask and extracts the ice lake through the first student model trained in the ice lake extraction student network training module, and then inputs the extracted ice lake image into the second student model trained in the water ice classification student network training module to obtain the water ice classification result.

[0033] For other contents of this embodiment, please refer to Embodiment 1.

[0034] In summary, the water ice classification method and system for multispectral images based on dual knowledge distillation designed by the present invention are used to achieve high-efficiency and high-quality multispectral remote sensing image water ice classification tasks, and can be applied to various engineering application scenarios with limited resources.

[0035] It should be noted that the above description is only a detailed description of the preferred embodiments and principles of the present invention. For ordinary technicians in this field, there will be changes in the specific implementation methods based on the ideas provided by the present invention, and these changes should also be regarded as the scope of protection of the present invention.

Claims

1. A water ice classification method for multispectral remote sensing images based on dual knowledge distillation, characterized in that: The steps include: Step 1, acquiring multispectral remote sensing images and performing data preprocessing; Step 2: Build the Ice Lake Extraction Dataless Knowledge Distillation Network, take the Ice Lake Index ILI as the first teacher model, build the U-net network as the first student model, and complete the dataless knowledge distillation; Step 3, constructing an Ice Lake Extraction Dataless Knowledge Distillation Framework, inputting the data preprocessed in Step 1 as a training set into the Ice Lake Extraction Dataless Knowledge Distillation Network described in Step 2, outputting the binary masks of Ice Lake Extraction respectively, and performing DICE loss calculation on the Ice Lake mask output of the first teacher model ILI and the Ice Lake mask output of the first student model U-net as the supervision loss; Step 4: construct a water ice classification teacher-student offline distillation network to classify water ice in images after ice lake extraction; the second teacher model uses a trained U-net classification model, and the second student model uses a lightweight model; Step 5, construct a knowledge distillation framework for water ice classification, perform ice lake extraction on the image through the first student model trained in step 3, and use the output of the first student model as the input of the water ice classification teacher-student offline distillation network in step 4 to train water ice classification; wherein the second teacher model and the second student model in the water ice classification teacher-student offline distillation network respectively output water ice classification results; In step 6, the remote sensing image is used as input, and ice lake extraction is performed through the ice lake extraction data-free knowledge distillation framework in step 3, and then water ice classification is performed through the water ice classification teacher-student offline distillation framework in step 5.

2. The water ice classification method for multispectral remote sensing images based on dual knowledge distillation according to claim 1, characterized in that: Step 1 includes the following steps: S11, grouping the multispectral remote sensing images by band according to spatial resolution, and reconstructing the multispectral remote sensing images with different spatial resolutions by super-resolution, and uniformly processing them to the highest spatial resolution; S12, cutting the image reconstructed in step S11 to form a training set.

3. The water ice classification method for multispectral remote sensing images based on dual knowledge distillation according to claim 1, characterized in that: In step 2, the U-net network includes an encoder, a decoder and an output layer; the number of parameters of the first student model is smaller than the number of parameters of the first teacher model.

4. The water ice classification method for multispectral remote sensing images based on dual knowledge distillation according to claim 1, characterized in that: In step 3, the supervised loss is used as the network training error to update the first student model for iterative training, so as to maximize the learning of the feature knowledge of the first teacher model by the first student model, thereby achieving data-free knowledge distillation; the supervised loss is calculated using the DICE function and used for back propagation to guide the training of the first student model, which is specifically expressed as: ； in, is the loss function of the Ice Lake extraction network, M represents the total number of pixels in the image, is the sum of the probabilities of the predicted segmentation results at pixel i; is the binary value of the true label at pixel i.

5. The water ice classification method for multispectral remote sensing images based on dual knowledge distillation according to claim 1, characterized in that: In step 4, the second teacher model includes an encoder, a decoder and an output layer; the second student model has the same architecture as the second teacher model, and the number of parameters of the second student model is smaller than the number of parameters of the second teacher model.

6. The water ice classification method for multispectral remote sensing images based on dual knowledge distillation according to claim 1, characterized in that: Step 5 also includes the following steps: S51, through Fourier transformation, the domain-invariant knowledge and domain-specific knowledge of the second teacher model are separated. When training the teacher-student model, the second student model acquires the domain-invariant knowledge by learning the second teacher model and acquires the domain-specific knowledge through labels; S52, the output of the second teacher model is compared with the output of the second student model as the teacher supervision loss, and the data supervision is performed by manually annotating the generated label data and comparing it with the second student model as the data supervision loss; The teacher supervision loss and data supervision loss are both calculated using the DICE function and used for back propagation to guide the training of the second student model, specifically expressed as: ； ； ； in, is the loss function for water ice classification, is the data supervision loss, is the teacher supervision loss; N represents the total number of pixels in the image; is the sum of the probabilities of the second student model predicting the segmentation result at pixel i; is the sum of the probabilities of the second teacher model predicting the segmentation result at pixel i, is the binary value of the true label segmentation result at pixel i; and They are weight coefficients for balancing losses in different aspects.

7. A water ice classification system for multispectral remote sensing images based on dual knowledge distillation, based on the method described in any one of claims 1 to 6, characterized in that: Includes the following modules: Image acquisition and preprocessing module, which acquires multispectral remote sensing images and performs data preprocessing; Ice Lake extracts data-free teacher-student offline distillation network construction modules, builds a teacher-student offline distillation network based on multispectral remote sensing images, takes the Ice Lake Index ILI as the first teacher model, and builds a U-net network as the first student model; The loss calculation module builds a knowledge distillation framework for Ice Lake Extraction for multispectral remote sensing images, inputs the training set in the image acquisition and preprocessing module into the first teacher model and the first student model in the Ice Lake Extraction Dataless Teacher-Student Offline Distillation Network Construction Module, outputs the Ice Lake mask respectively, and calculates the loss of the outputs of the first teacher model and the first student model as the supervision loss; The water ice classification teacher-student offline distillation network construction module builds a water ice classification teacher-student offline distillation network based on multispectral remote sensing images, classifies water ice in images after ice lake extraction, uses the trained U-net classification model as the second teacher model, and uses the lightweight model as the second student model; The total loss module builds a knowledge distillation framework for water ice classification for multispectral remote sensing images. The image is subjected to ice lake extraction through the first student model in the trained ice lake extraction data-free knowledge distillation framework, and the output of the first student model is used as the input of the water ice classification teacher-student offline distillation network for water ice classification training. The teacher model and the student model of the water ice classification teacher-student offline distillation network output the water ice classification results respectively, and the output of the second teacher model and the second student model is used for loss calculation as the teacher supervision loss, i.e., the knowledge distillation loss. The super-resolution remote sensing image output by the second student model and the training set label are subjected to loss calculation as the data supervision loss; the knowledge distillation loss and the data supervision loss are weighted and added to calculate the total loss; The water ice classification module takes the multispectral remote sensing image as input, obtains the ice lake mask and extracts the ice lake through the first student model trained in the ice lake extraction student network training module, and then inputs the extracted ice lake image into the second student model trained in the water ice classification student network training module to obtain the water ice classification result.

Images